基于双向循环神经网络的汉语语音识别*

当前基于深度神经网络模型中,虽然其隐含层可设置多层,对复杂问题适应能力强,但每层之间的节点连接是相互独立的,这种结构特性导致了在语音序列中无法利用上下文相关信息来提高识别效果,而传统的循环神经网络虽然做出了改进,但是只能对上文信息进行利用。针对以上问题,该文采用可以同时利用语音序列中上下文相关信息的双向循环神经网络模型与深度神经网络模型相结合,并应用于语音识别。构建具有5层隐含层的模型,其中第3层为双向循环神经网络结构,其他层采用深度神经网络结构。实验结果表明:加入了双向循环神经网络结构的模型与其他模型相比,较好地提高了识别正确率;噪声对双向循环神经网络汉语识别有重要影响,尤其是训练集和测试集附加噪声类型不同时,单一的含噪声语音的训练模型无法适应不同噪声类型的语音识别;调整神经网络模型中隐含层神经元数量后,识别正确率并不是一直随着隐含层中神经元数量的增加而增加,神经元数量数目增加到一定程度后正确率出现了降低的趋势。     

SE-MCNN-CTC的中文语音识别声学模型

针对国内外缺少对振动轮噪声预估的问题,以某型振动轮为研究对象,首先基于动力学有限元理论对振动轮进行频率响应分析,其次采用声学边界元技术对振动轮辐射噪声进行了数值模拟,并通过实验验证了仿真结果的准确性,然后比较了垂直振动与圆周振动两种不同激振形式对辐射噪声的影响,得出垂直振动辐射噪声低的结论,最后对驾驶室声腔模态进行了仿真,与振动轮激振频率相近发生共振。通过调整激振频率,降低了司机耳旁噪声。所得研究成果可为振动轮辐射噪声的预估与改进提供一种切实可行的参考依据。     

Aerodynamic and Acoustic Characteristics of Voice in Prader-Willi Syndrome

Prader-Willi syndrome (PWS) is a multisystem disorder caused by DNA abnormalities involving chromosome 15. Major characteristics are infant hypotonia, hypogonadism, mental retardation, a short stature, atypical facial appearance, and the onset of obesity due to insatiable hunger in early childhood. Also, speech and language abnormalities have been reported including voice disorders. These have seldom been studied in detail, however. This paper reports the results of an acoustic and aerodynamic investigation of the voice in 22 individuals with PWS. Two age groups were distinguished, a group of children [chronological age (CA) 6 years, 7 months through 11 years, 7 months; total intelligence quotient (TIQ) 40-88] and a group of adolescents and adults (CA 17 years, 1 month through 29 years, 5 months; TIQ 41-94). Both aerodynamic and acoustic parameters were obtained and compared with normative data from the Belgian Study Group on Voice Disorders. It was found that voice difficulties do commonly occur in individuals with PWS including impairment of frequency levels, voice quality, and poor aerodynamic capabilities.     

基于双传声器的蓝牙耳机降噪算法

用于免提通信设备的语音增强算法一直是研究的热点问题,而算法处理结果的音质问题近年来也备受关注。针对基于双传声器降噪的蓝牙耳机系统,将常用多通道传声器降噪算法归纳为基于相干函数法和基于空间预分离法这两大类进行分析和比较。基于相干函数法利用两个通道间信号的相干函数对含噪信号滤波达到降噪目的,而基于空间预分离法利用空间特性从含噪信号中分离出噪声参考信号来消除噪声。分析基于降噪量、语音音质和综合性能三个指标,从约束语音损伤的角度分析最优解的形式,并对比两类算法的实际性能。结果表明选择合适的算法可权衡降噪量与语音损伤,达到较好的综合性能。     

Accuracy of speech transmission index predictions based on the reverberation time and signal-to-noise ratio

This paper examines the accuracy of the speech transmission index (STI) calculated from the reverberation time (T) and signal-to-noise ratio (L_SN) of enclosed spaces. Differences between measured and predicted STIs have been analysed in two rooms (reverberant vs. absorbent), for a wide range of absorption conditions and signal-to-noise ratios (sixteen tests). The STI was measured using maximum length sequence analysis and predictions were calculated using either measured or predicted values of T and L_SN, the latter assuming diffuse sound field conditions. The results obtained for all the conditions tested showed that STI predictions based on T and L_SN tend to underestimate the STI, with differences between measured and predicted STIs always lower than 0.1 (on a 0.0–1.0 scale), and on average lower than 0.06. According to previous research, these differences are noticeable and therefore non-negligible, as 0.03 is the just noticeable difference in STI. The use of either measured or predicted values of T and L_SN provided similar STI predictions (i.e. non-noticeable changes), with differences between predictions that are on average lower than 0.03 for the absorbent room, and lower than 0.01 for the reverberant room.     

Simulating the effect of acoustic treatment types for residential balconies with road traffic noise

Development of design guides to estimate the difference in speech interference level due to road traffic noise between a reference position and balcony position or façade position is explored. A previously established and validated theoretical model incorporating direct, specular and diffuse reflection paths is used to create a database of results across a large number of scenarios. Nine balcony types with variable acoustic treatments are assessed to provide acoustic design guidance on optimised selection of balcony acoustic treatments based on location and street type. In total, the results database contains 9720 scenarios on which multivariate linear regression is conducted in order to derive an appropriate design guide equation. The best fit regression derived is a multivariable linear equation including modified exponential equations on each of nine deciding variables, (1) diffraction path difference, (2) ratio of total specular energy to direct energy, (3) distance loss between reference position and receiver position, (4) distance from source to balcony façade, (5) height of balcony floor above street, (6) balcony depth, (7) height of opposite buildings, (8) diffusion coefficient of buildings and (9) balcony average absorption. Overall, the regression correlation coefficient, R², is 0.89 with 95% confidence standard error of ±3.4 dB.